| 1 | |
| 2 | |
| 3 | * mm/readahead.c - address_space-level file readahead.
|
| 4 | *
|
| 5 | * Copyright (C) 2002, Linus Torvalds
|
| 6 | *
|
| 7 | * 09Apr2002 Andrew Morton
|
| 8 | * Initial version.
|
| 9 |
|
| 10 |
|
| 11 | |
| 12 | * DOC: Readahead Overview
|
| 13 | *
|
| 14 | * Readahead is used to read content into the page cache before it is
|
| 15 | * explicitly requested by the application. Readahead only ever
|
| 16 | * attempts to read folios that are not yet in the page cache. If a
|
| 17 | * folio is present but not up-to-date, readahead will not try to read
|
| 18 | * it. In that case a simple ->read_folio() will be requested.
|
| 19 | *
|
| 20 | * Readahead is triggered when an application read request (whether a
|
| 21 | * system call or a page fault) finds that the requested folio is not in
|
| 22 | * the page cache, or that it is in the page cache and has the
|
| 23 | * readahead flag set. This flag indicates that the folio was read
|
| 24 | * as part of a previous readahead request and now that it has been
|
| 25 | * accessed, it is time for the next readahead.
|
| 26 | *
|
| 27 | * Each readahead request is partly synchronous read, and partly async
|
| 28 | * readahead. This is reflected in the struct file_ra_state which
|
| 29 | * contains ->size being the total number of pages, and ->async_size
|
| 30 | * which is the number of pages in the async section. The readahead
|
| 31 | * flag will be set on the first folio in this async section to trigger
|
| 32 | * a subsequent readahead. Once a series of sequential reads has been
|
| 33 | * established, there should be no need for a synchronous component and
|
| 34 | * all readahead request will be fully asynchronous.
|
| 35 | *
|
| 36 | * When either of the triggers causes a readahead, three numbers need
|
| 37 | * to be determined: the start of the region to read, the size of the
|
| 38 | * region, and the size of the async tail.
|
| 39 | *
|
| 40 | * The start of the region is simply the first page address at or after
|
| 41 | * the accessed address, which is not currently populated in the page
|
| 42 | * cache. This is found with a simple search in the page cache.
|
| 43 | *
|
| 44 | * The size of the async tail is determined by subtracting the size that
|
| 45 | * was explicitly requested from the determined request size, unless
|
| 46 | * this would be less than zero - then zero is used. NOTE THIS
|
| 47 | * CALCULATION IS WRONG WHEN THE START OF THE REGION IS NOT THE ACCESSED
|
| 48 | * PAGE. ALSO THIS CALCULATION IS NOT USED CONSISTENTLY.
|
| 49 | *
|
| 50 | * The size of the region is normally determined from the size of the
|
| 51 | * previous readahead which loaded the preceding pages. This may be
|
| 52 | * discovered from the struct file_ra_state for simple sequential reads,
|
| 53 | * or from examining the state of the page cache when multiple
|
| 54 | * sequential reads are interleaved. Specifically: where the readahead
|
| 55 | * was triggered by the readahead flag, the size of the previous
|
| 56 | * readahead is assumed to be the number of pages from the triggering
|
| 57 | * page to the start of the new readahead. In these cases, the size of
|
| 58 | * the previous readahead is scaled, often doubled, for the new
|
| 59 | * readahead, though see get_next_ra_size() for details.
|
| 60 | *
|
| 61 | * If the size of the previous read cannot be determined, the number of
|
| 62 | * preceding pages in the page cache is used to estimate the size of
|
| 63 | * a previous read. This estimate could easily be misled by random
|
| 64 | * reads being coincidentally adjacent, so it is ignored unless it is
|
| 65 | * larger than the current request, and it is not scaled up, unless it
|
| 66 | * is at the start of file.
|
| 67 | *
|
| 68 | * In general readahead is accelerated at the start of the file, as
|
| 69 | * reads from there are often sequential. There are other minor
|
| 70 | * adjustments to the readahead size in various special cases and these
|
| 71 | * are best discovered by reading the code.
|
| 72 | *
|
| 73 | * The above calculation, based on the previous readahead size,
|
| 74 | * determines the size of the readahead, to which any requested read
|
| 75 | * size may be added.
|
| 76 | *
|
| 77 | * Readahead requests are sent to the filesystem using the ->readahead()
|
| 78 | * address space operation, for which mpage_readahead() is a canonical
|
| 79 | * implementation. ->readahead() should normally initiate reads on all
|
| 80 | * folios, but may fail to read any or all folios without causing an I/O
|
| 81 | * error. The page cache reading code will issue a ->read_folio() request
|
| 82 | * for any folio which ->readahead() did not read, and only an error
|
| 83 | * from this will be final.
|
| 84 | *
|
| 85 | * ->readahead() will generally call readahead_folio() repeatedly to get
|
| 86 | * each folio from those prepared for readahead. It may fail to read a
|
| 87 | * folio by:
|
| 88 | *
|
| 89 | * * not calling readahead_folio() sufficiently many times, effectively
|
| 90 | * ignoring some folios, as might be appropriate if the path to
|
| 91 | * storage is congested.
|
| 92 | *
|
| 93 | * * failing to actually submit a read request for a given folio,
|
| 94 | * possibly due to insufficient resources, or
|
| 95 | *
|
| 96 | * * getting an error during subsequent processing of a request.
|
| 97 | *
|
| 98 | * In the last two cases, the folio should be unlocked by the filesystem
|
| 99 | * to indicate that the read attempt has failed. In the first case the
|
| 100 | * folio will be unlocked by the VFS.
|
| 101 | *
|
| 102 | * Those folios not in the final ``async_size`` of the request should be
|
| 103 | * considered to be important and ->readahead() should not fail them due
|
| 104 | * to congestion or temporary resource unavailability, but should wait
|
| 105 | * for necessary resources (e.g. memory or indexing information) to
|
| 106 | * become available. Folios in the final ``async_size`` may be
|
| 107 | * considered less urgent and failure to read them is more acceptable.
|
| 108 | * In this case it is best to use filemap_remove_folio() to remove the
|
| 109 | * folios from the page cache as is automatically done for folios that
|
| 110 | * were not fetched with readahead_folio(). This will allow a
|
| 111 | * subsequent synchronous readahead request to try them again. If they
|
| 112 | * are left in the page cache, then they will be read individually using
|
| 113 | * ->read_folio() which may be less efficient.
|
| 114 |
|
| 115 |
|
| 116 | #include <linux/blk_plug.h>
|
| 117 | #include <linux/kernel.h>
|
| 118 | #include <linux/dax.h>
|
| 119 | #include <linux/gfp.h>
|
| 120 | #include <linux/export.h>
|
| 121 | #include <linux/backing-dev.h>
|
| 122 | #include <linux/task_io_accounting_ops.h>
|
| 123 | #include <linux/pagemap.h>
|
| 124 | #include <linux/psi.h>
|
| 125 | #include <linux/syscalls.h>
|
| 126 | #include <linux/file.h>
|
| 127 | #include <linux/mm_inline.h>
|
| 128 | #include <linux/blk-cgroup.h>
|
| 129 | #include <linux/fadvise.h>
|
| 130 | #include <linux/sched/mm.h>
|
| 131 |
|
| 132 | #define CREATE_TRACE_POINTS
|
| 133 | #include <trace/events/readahead.h>
|
| 134 |
|
| 135 | #include "internal.h"
|
| 136 |
|
| 137 | |
| 138 | * Initialise a struct file's readahead state. Assumes that the caller has
|
| 139 | * memset *ra to zero.
|
| 140 |
|
| 141 | void
|
| 142 | file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
|
| 143 | {
|
| 144 | ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
|
| 145 | ra->prev_pos = -1;
|
| 146 | }
|
| 147 | EXPORT_SYMBOL_GPL(file_ra_state_init);
|
| 148 |
|
| 149 | |
| 150 | * read_pages() - Start IO for a contiguous range of allocated folios in the
|
| 151 | * page cache.
|
| 152 | * @rac: Readahead control.
|
| 153 | *
|
| 154 | * When read_pages() returns, it is guaranteed that all of the folios will have
|
| 155 | * been processed or removed so that ``readahead_count(rac) == 0``. However,
|
| 156 | * that does not imply that ``readahead_index(rac)`` will be updated to point
|
| 157 | * to the end of the originally requested range because, for example, the
|
| 158 | * filesystem may expand the range upwards.
|
| 159 |
|
| 160 | static void read_pages(struct readahead_control *rac)
|
| 161 | {
|
| 162 | const struct address_space_operations *aops = rac->mapping->a_ops;
|
| 163 | struct folio *folio;
|
| 164 | struct blk_plug plug;
|
| 165 |
|
| 166 | if (!readahead_count(rac))
|
| 167 | return;
|
| 168 |
|
| 169 | if (unlikely(rac->_workingset))
|
| 170 | psi_memstall_enter(&rac->_pflags);
|
| 171 | blk_start_plug(&plug);
|
| 172 |
|
| 173 | if (aops->readahead) {
|
| 174 | aops->readahead(rac);
|
| 175 |
|
| 176 | while ((folio = readahead_folio(rac)) != NULL) {
|
| 177 | folio_get(folio);
|
| 178 | filemap_remove_folio(folio);
|
| 179 | folio_unlock(folio);
|
| 180 | folio_put(folio);
|
| 181 | }
|
| 182 | } else {
|
| 183 | while ((folio = readahead_folio(rac)) != NULL)
|
| 184 | aops->read_folio(rac->file, folio);
|
| 185 | }
|
| 186 |
|
| 187 | blk_finish_plug(&plug);
|
| 188 | if (unlikely(rac->_workingset))
|
| 189 | psi_memstall_leave(&rac->_pflags);
|
| 190 | rac->_workingset = false;
|
| 191 |
|
| 192 | BUG_ON(readahead_count(rac));
|
| 193 | }
|
| 194 |
|
| 195 | static struct folio *ractl_alloc_folio(struct readahead_control *ractl,
|
| 196 | gfp_t gfp_mask, unsigned int order)
|
| 197 | {
|
| 198 | struct folio *folio;
|
| 199 |
|
| 200 | folio = filemap_alloc_folio(gfp_mask, order, NULL);
|
| 201 | if (folio && ractl->dropbehind)
|
| 202 | __folio_set_dropbehind(folio);
|
| 203 |
|
| 204 | return folio;
|
| 205 | }
|
| 206 |
|
| 207 | |
| 208 | * page_cache_ra_unbounded - Start unchecked readahead.
|
| 209 | * @ractl: Readahead control.
|
| 210 | * @nr_to_read: The number of pages to read.
|
| 211 | * @lookahead_size: Where to start the next readahead.
|
| 212 | *
|
| 213 | * This function is for filesystems to call when they want to start
|
| 214 | * readahead beyond a file's stated i_size. This is almost certainly
|
| 215 | * not the function you want to call. Use page_cache_async_readahead()
|
| 216 | * or page_cache_sync_readahead() instead.
|
| 217 | *
|
| 218 | * Context: File is referenced by caller, and ractl->mapping->invalidate_lock
|
| 219 | * must be held by the caller at least in shared mode. Mutexes may be held by
|
| 220 | * caller. May sleep, but will not reenter filesystem to reclaim memory.
|
| 221 |
|
| 222 | void page_cache_ra_unbounded(struct readahead_control *ractl,
|
| 223 | unsigned long nr_to_read, unsigned long lookahead_size)
|
| 224 | {
|
| 225 | struct address_space *mapping = ractl->mapping;
|
| 226 | unsigned long index = readahead_index(ractl);
|
| 227 | gfp_t gfp_mask = readahead_gfp_mask(mapping);
|
| 228 | unsigned long mark = ULONG_MAX, i = 0;
|
| 229 | unsigned int min_nrpages = mapping_min_folio_nrpages(mapping);
|
| 230 |
|
| 231 | |
| 232 | * Partway through the readahead operation, we will have added
|
| 233 | * locked pages to the page cache, but will not yet have submitted
|
| 234 | * them for I/O. Adding another page may need to allocate memory,
|
| 235 | * which can trigger memory reclaim. Telling the VM we're in
|
| 236 | * the middle of a filesystem operation will cause it to not
|
| 237 | * touch file-backed pages, preventing a deadlock. Most (all?)
|
| 238 | * filesystems already specify __GFP_NOFS in their mapping's
|
| 239 | * gfp_mask, but let's be explicit here.
|
| 240 |
|
| 241 | unsigned int nofs = memalloc_nofs_save();
|
| 242 |
|
| 243 | lockdep_assert_held(&mapping->invalidate_lock);
|
| 244 |
|
| 245 | trace_page_cache_ra_unbounded(mapping->host, index, nr_to_read,
|
| 246 | lookahead_size);
|
| 247 | index = mapping_align_index(mapping, index);
|
| 248 |
|
| 249 | |
| 250 | * As iterator `i` is aligned to min_nrpages, round_up the
|
| 251 | * difference between nr_to_read and lookahead_size to mark the
|
| 252 | * index that only has lookahead or "async_region" to set the
|
| 253 | * readahead flag.
|
| 254 |
|
| 255 | if (lookahead_size <= nr_to_read) {
|
| 256 | unsigned long ra_folio_index;
|
| 257 |
|
| 258 | ra_folio_index = round_up(readahead_index(ractl) +
|
| 259 | nr_to_read - lookahead_size,
|
| 260 | min_nrpages);
|
| 261 | mark = ra_folio_index - index;
|
| 262 | }
|
| 263 | nr_to_read += readahead_index(ractl) - index;
|
| 264 | ractl->_index = index;
|
| 265 |
|
| 266 | |
| 267 | * Preallocate as many pages as we will need.
|
| 268 |
|
| 269 | while (i < nr_to_read) {
|
| 270 | struct folio *folio = xa_load(&mapping->i_pages, index + i);
|
| 271 | int ret;
|
| 272 |
|
| 273 | if (folio && !xa_is_value(folio)) {
|
| 274 | |
| 275 | * Page already present? Kick off the current batch
|
| 276 | * of contiguous pages before continuing with the
|
| 277 | * next batch. This page may be the one we would
|
| 278 | * have intended to mark as Readahead, but we don't
|
| 279 | * have a stable reference to this page, and it's
|
| 280 | * not worth getting one just for that.
|
| 281 |
|
| 282 | read_pages(ractl);
|
| 283 | ractl->_index += min_nrpages;
|
| 284 | i = ractl->_index - index;
|
| 285 | continue;
|
| 286 | }
|
| 287 |
|
| 288 | folio = ractl_alloc_folio(ractl, gfp_mask,
|
| 289 | mapping_min_folio_order(mapping));
|
| 290 | if (!folio)
|
| 291 | break;
|
| 292 |
|
| 293 | ret = filemap_add_folio(mapping, folio, index + i, gfp_mask);
|
| 294 | if (ret < 0) {
|
| 295 | folio_put(folio);
|
| 296 | if (ret == -ENOMEM)
|
| 297 | break;
|
| 298 | read_pages(ractl);
|
| 299 | ractl->_index += min_nrpages;
|
| 300 | i = ractl->_index - index;
|
| 301 | continue;
|
| 302 | }
|
| 303 | if (i == mark)
|
| 304 | folio_set_readahead(folio);
|
| 305 | ractl->_workingset |= folio_test_workingset(folio);
|
| 306 | ractl->_nr_pages += min_nrpages;
|
| 307 | i += min_nrpages;
|
| 308 | }
|
| 309 |
|
| 310 | |
| 311 | * Now start the IO. We ignore I/O errors - if the folio is not
|
| 312 | * uptodate then the caller will launch read_folio again, and
|
| 313 | * will then handle the error.
|
| 314 |
|
| 315 | read_pages(ractl);
|
| 316 | memalloc_nofs_restore(nofs);
|
| 317 | }
|
| 318 | EXPORT_SYMBOL_GPL(page_cache_ra_unbounded);
|
| 319 |
|
| 320 | |
| 321 | * do_page_cache_ra() actually reads a chunk of disk. It allocates
|
| 322 | * the pages first, then submits them for I/O. This avoids the very bad
|
| 323 | * behaviour which would occur if page allocations are causing VM writeback.
|
| 324 | * We really don't want to intermingle reads and writes like that.
|
| 325 |
|
| 326 | static void do_page_cache_ra(struct readahead_control *ractl,
|
| 327 | unsigned long nr_to_read, unsigned long lookahead_size)
|
| 328 | {
|
| 329 | struct address_space *mapping = ractl->mapping;
|
| 330 | unsigned long index = readahead_index(ractl);
|
| 331 | loff_t isize = i_size_read(mapping->host);
|
| 332 | pgoff_t end_index;
|
| 333 |
|
| 334 | if (isize == 0)
|
| 335 | return;
|
| 336 |
|
| 337 | end_index = (isize - 1) >> PAGE_SHIFT;
|
| 338 | if (index > end_index)
|
| 339 | return;
|
| 340 |
|
| 341 | if (nr_to_read > end_index - index) {
|
| 342 | nr_to_read = end_index - index + 1;
|
| 343 |
|
| 344 | lookahead_size = 0;
|
| 345 | }
|
| 346 |
|
| 347 | filemap_invalidate_lock_shared(mapping);
|
| 348 | page_cache_ra_unbounded(ractl, nr_to_read, lookahead_size);
|
| 349 | filemap_invalidate_unlock_shared(mapping);
|
| 350 | }
|
| 351 |
|
| 352 | |
| 353 | * Chunk the readahead into 2 megabyte units, so that we don't pin too much
|
| 354 | * memory at once.
|
| 355 |
|
| 356 | void force_page_cache_ra(struct readahead_control *ractl,
|
| 357 | unsigned long nr_to_read)
|
| 358 | {
|
| 359 | struct address_space *mapping = ractl->mapping;
|
| 360 | struct file_ra_state *ra = ractl->ra;
|
| 361 | struct backing_dev_info *bdi = inode_to_bdi(mapping->host);
|
| 362 | unsigned long max_pages;
|
| 363 |
|
| 364 | if (unlikely(!mapping->a_ops->read_folio && !mapping->a_ops->readahead))
|
| 365 | return;
|
| 366 |
|
| 367 | |
| 368 | * If the request exceeds the readahead window, allow the read to
|
| 369 | * be up to the optimal hardware IO size
|
| 370 |
|
| 371 | max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
|
| 372 | nr_to_read = min_t(unsigned long, nr_to_read, max_pages);
|
| 373 | while (nr_to_read) {
|
| 374 | unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_SIZE;
|
| 375 |
|
| 376 | if (this_chunk > nr_to_read)
|
| 377 | this_chunk = nr_to_read;
|
| 378 | do_page_cache_ra(ractl, this_chunk, 0);
|
| 379 |
|
| 380 | nr_to_read -= this_chunk;
|
| 381 | }
|
| 382 | }
|
| 383 |
|
| 384 | |
| 385 | * Set the initial window size, round to next power of 2 and square
|
| 386 | * for small size, x 4 for medium, and x 2 for large
|
| 387 | * for 128k (32 page) max ra
|
| 388 | * 1-2 page = 16k, 3-4 page 32k, 5-8 page = 64k, > 8 page = 128k initial
|
| 389 |
|
| 390 | static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
|
| 391 | {
|
| 392 | unsigned long newsize = roundup_pow_of_two(size);
|
| 393 |
|
| 394 | if (newsize <= max / 32)
|
| 395 | newsize = newsize * 4;
|
| 396 | else if (newsize <= max / 4)
|
| 397 | newsize = newsize * 2;
|
| 398 | else
|
| 399 | newsize = max;
|
| 400 |
|
| 401 | return newsize;
|
| 402 | }
|
| 403 |
|
| 404 | |
| 405 | * Get the previous window size, ramp it up, and
|
| 406 | * return it as the new window size.
|
| 407 |
|
| 408 | static unsigned long get_next_ra_size(struct file_ra_state *ra,
|
| 409 | unsigned long max)
|
| 410 | {
|
| 411 | unsigned long cur = ra->size;
|
| 412 |
|
| 413 | if (cur < max / 16)
|
| 414 | return 4 * cur;
|
| 415 | if (cur <= max / 2)
|
| 416 | return 2 * cur;
|
| 417 | return max;
|
| 418 | }
|
| 419 |
|
| 420 | |
| 421 | * On-demand readahead design.
|
| 422 | *
|
| 423 | * The fields in struct file_ra_state represent the most-recently-executed
|
| 424 | * readahead attempt:
|
| 425 | *
|
| 426 | * |<----- async_size ---------|
|
| 427 | * |------------------- size -------------------->|
|
| 428 | * |==================#===========================|
|
| 429 | * ^start ^page marked with PG_readahead
|
| 430 | *
|
| 431 | * To overlap application thinking time and disk I/O time, we do
|
| 432 | * `readahead pipelining': Do not wait until the application consumed all
|
| 433 | * readahead pages and stalled on the missing page at readahead_index;
|
| 434 | * Instead, submit an asynchronous readahead I/O as soon as there are
|
| 435 | * only async_size pages left in the readahead window. Normally async_size
|
| 436 | * will be equal to size, for maximum pipelining.
|
| 437 | *
|
| 438 | * In interleaved sequential reads, concurrent streams on the same fd can
|
| 439 | * be invalidating each other's readahead state. So we flag the new readahead
|
| 440 | * page at (start+size-async_size) with PG_readahead, and use it as readahead
|
| 441 | * indicator. The flag won't be set on already cached pages, to avoid the
|
| 442 | * readahead-for-nothing fuss, saving pointless page cache lookups.
|
| 443 | *
|
| 444 | * prev_pos tracks the last visited byte in the _previous_ read request.
|
| 445 | * It should be maintained by the caller, and will be used for detecting
|
| 446 | * small random reads. Note that the readahead algorithm checks loosely
|
| 447 | * for sequential patterns. Hence interleaved reads might be served as
|
| 448 | * sequential ones.
|
| 449 | *
|
| 450 | * There is a special-case: if the first page which the application tries to
|
| 451 | * read happens to be the first page of the file, it is assumed that a linear
|
| 452 | * read is about to happen and the window is immediately set to the initial size
|
| 453 | * based on I/O request size and the max_readahead.
|
| 454 | *
|
| 455 | * The code ramps up the readahead size aggressively at first, but slow down as
|
| 456 | * it approaches max_readahead.
|
| 457 |
|
| 458 |
|
| 459 | static inline int ra_alloc_folio(struct readahead_control *ractl, pgoff_t index,
|
| 460 | pgoff_t mark, unsigned int order, gfp_t gfp)
|
| 461 | {
|
| 462 | int err;
|
| 463 | struct folio *folio = ractl_alloc_folio(ractl, gfp, order);
|
| 464 |
|
| 465 | if (!folio)
|
| 466 | return -ENOMEM;
|
| 467 | mark = round_down(mark, 1UL << order);
|
| 468 | if (index == mark)
|
| 469 | folio_set_readahead(folio);
|
| 470 | err = filemap_add_folio(ractl->mapping, folio, index, gfp);
|
| 471 | if (err) {
|
| 472 | folio_put(folio);
|
| 473 | return err;
|
| 474 | }
|
| 475 |
|
| 476 | ractl->_nr_pages += 1UL << order;
|
| 477 | ractl->_workingset |= folio_test_workingset(folio);
|
| 478 | return 0;
|
| 479 | }
|
| 480 |
|
| 481 | void page_cache_ra_order(struct readahead_control *ractl,
|
| 482 | struct file_ra_state *ra)
|
| 483 | {
|
| 484 | struct address_space *mapping = ractl->mapping;
|
| 485 | pgoff_t start = readahead_index(ractl);
|
| 486 | pgoff_t index = start;
|
| 487 | unsigned int min_order = mapping_min_folio_order(mapping);
|
| 488 | pgoff_t limit = (i_size_read(mapping->host) - 1) >> PAGE_SHIFT;
|
| 489 | pgoff_t mark;
|
| 490 | unsigned int nofs;
|
| 491 | int err = 0;
|
| 492 | gfp_t gfp = readahead_gfp_mask(mapping);
|
| 493 | unsigned int new_order = ra->order;
|
| 494 |
|
| 495 | trace_page_cache_ra_order(mapping->host, start, ra);
|
| 496 | if (!mapping_large_folio_support(mapping)) {
|
| 497 | ra->order = 0;
|
| 498 | goto fallback;
|
| 499 | }
|
| 500 |
|
| 501 | if (limit > index + ra->size - 1) {
|
| 502 | limit = index + ra->size - 1;
|
| 503 | mark = index + ra->size - ra->async_size;
|
| 504 | } else {
|
| 505 |
|
| 506 | mark = ULONG_MAX;
|
| 507 | }
|
| 508 |
|
| 509 | new_order = min(mapping_max_folio_order(mapping), new_order);
|
| 510 | new_order = min_t(unsigned int, new_order, ilog2(ra->size));
|
| 511 | new_order = max(new_order, min_order);
|
| 512 |
|
| 513 | ra->order = new_order;
|
| 514 |
|
| 515 |
|
| 516 | nofs = memalloc_nofs_save();
|
| 517 | filemap_invalidate_lock_shared(mapping);
|
| 518 | |
| 519 | * If the new_order is greater than min_order and index is
|
| 520 | * already aligned to new_order, then this will be noop as index
|
| 521 | * aligned to new_order should also be aligned to min_order.
|
| 522 |
|
| 523 | ractl->_index = mapping_align_index(mapping, index);
|
| 524 | index = readahead_index(ractl);
|
| 525 |
|
| 526 | while (index <= limit) {
|
| 527 | unsigned int order = new_order;
|
| 528 |
|
| 529 |
|
| 530 | if (index & ((1UL << order) - 1))
|
| 531 | order = __ffs(index);
|
| 532 |
|
| 533 | while (order > min_order && index + (1UL << order) - 1 > limit)
|
| 534 | order--;
|
| 535 | err = ra_alloc_folio(ractl, index, mark, order, gfp);
|
| 536 | if (err)
|
| 537 | break;
|
| 538 | index += 1UL << order;
|
| 539 | }
|
| 540 |
|
| 541 | read_pages(ractl);
|
| 542 | filemap_invalidate_unlock_shared(mapping);
|
| 543 | memalloc_nofs_restore(nofs);
|
| 544 |
|
| 545 | |
| 546 | * If there were already pages in the page cache, then we may have
|
| 547 | * left some gaps. Let the regular readahead code take care of this
|
| 548 | * situation below.
|
| 549 |
|
| 550 | if (!err)
|
| 551 | return;
|
| 552 | fallback:
|
| 553 | |
| 554 | * ->readahead() may have updated readahead window size so we have to
|
| 555 | * check there's still something to read.
|
| 556 |
|
| 557 | if (ra->size > index - start)
|
| 558 | do_page_cache_ra(ractl, ra->size - (index - start),
|
| 559 | ra->async_size);
|
| 560 | }
|
| 561 |
|
| 562 | static unsigned long ractl_max_pages(struct readahead_control *ractl,
|
| 563 | unsigned long req_size)
|
| 564 | {
|
| 565 | struct backing_dev_info *bdi = inode_to_bdi(ractl->mapping->host);
|
| 566 | unsigned long max_pages = ractl->ra->ra_pages;
|
| 567 |
|
| 568 | |
| 569 | * If the request exceeds the readahead window, allow the read to
|
| 570 | * be up to the optimal hardware IO size
|
| 571 |
|
| 572 | if (req_size > max_pages && bdi->io_pages > max_pages)
|
| 573 | max_pages = min(req_size, bdi->io_pages);
|
| 574 | return max_pages;
|
| 575 | }
|
| 576 |
|
| 577 | void page_cache_sync_ra(struct readahead_control *ractl,
|
| 578 | unsigned long req_count)
|
| 579 | {
|
| 580 | pgoff_t index = readahead_index(ractl);
|
| 581 | bool do_forced_ra = ractl->file && (ractl->file->f_mode & FMODE_RANDOM);
|
| 582 | struct file_ra_state *ra = ractl->ra;
|
| 583 | unsigned long max_pages, contig_count;
|
| 584 | pgoff_t prev_index, miss;
|
| 585 |
|
| 586 | trace_page_cache_sync_ra(ractl->mapping->host, index, ra, req_count);
|
| 587 | |
| 588 | * Even if readahead is disabled, issue this request as readahead
|
| 589 | * as we'll need it to satisfy the requested range. The forced
|
| 590 | * readahead will do the right thing and limit the read to just the
|
| 591 | * requested range, which we'll set to 1 page for this case.
|
| 592 |
|
| 593 | if (!ra->ra_pages || blk_cgroup_congested()) {
|
| 594 | if (!ractl->file)
|
| 595 | return;
|
| 596 | req_count = 1;
|
| 597 | do_forced_ra = true;
|
| 598 | }
|
| 599 |
|
| 600 |
|
| 601 | if (do_forced_ra) {
|
| 602 | force_page_cache_ra(ractl, req_count);
|
| 603 | return;
|
| 604 | }
|
| 605 |
|
| 606 | max_pages = ractl_max_pages(ractl, req_count);
|
| 607 | prev_index = (unsigned long long)ra->prev_pos >> PAGE_SHIFT;
|
| 608 | |
| 609 | * A start of file, oversized read, or sequential cache miss:
|
| 610 | * trivial case: (index - prev_index) == 1
|
| 611 | * unaligned reads: (index - prev_index) == 0
|
| 612 |
|
| 613 | if (!index || req_count > max_pages || index - prev_index <= 1UL) {
|
| 614 | ra->start = index;
|
| 615 | ra->size = get_init_ra_size(req_count, max_pages);
|
| 616 | ra->async_size = ra->size > req_count ? ra->size - req_count :
|
| 617 | ra->size >> 1;
|
| 618 | goto readit;
|
| 619 | }
|
| 620 |
|
| 621 | |
| 622 | * Query the page cache and look for the traces(cached history pages)
|
| 623 | * that a sequential stream would leave behind.
|
| 624 |
|
| 625 | rcu_read_lock();
|
| 626 | miss = page_cache_prev_miss(ractl->mapping, index - 1, max_pages);
|
| 627 | rcu_read_unlock();
|
| 628 | contig_count = index - miss - 1;
|
| 629 | |
| 630 | * Standalone, small random read. Read as is, and do not pollute the
|
| 631 | * readahead state.
|
| 632 |
|
| 633 | if (contig_count <= req_count) {
|
| 634 | do_page_cache_ra(ractl, req_count, 0);
|
| 635 | return;
|
| 636 | }
|
| 637 | |
| 638 | * File cached from the beginning:
|
| 639 | * it is a strong indication of long-run stream (or whole-file-read)
|
| 640 |
|
| 641 | if (miss == ULONG_MAX)
|
| 642 | contig_count *= 2;
|
| 643 | ra->start = index;
|
| 644 | ra->size = min(contig_count + req_count, max_pages);
|
| 645 | ra->async_size = 1;
|
| 646 | readit:
|
| 647 | ra->order = 0;
|
| 648 | ractl->_index = ra->start;
|
| 649 | page_cache_ra_order(ractl, ra);
|
| 650 | }
|
| 651 | EXPORT_SYMBOL_GPL(page_cache_sync_ra);
|
| 652 |
|
| 653 | void page_cache_async_ra(struct readahead_control *ractl,
|
| 654 | struct folio *folio, unsigned long req_count)
|
| 655 | {
|
| 656 | unsigned long max_pages;
|
| 657 | struct file_ra_state *ra = ractl->ra;
|
| 658 | pgoff_t index = readahead_index(ractl);
|
| 659 | pgoff_t expected, start, end, aligned_end, align;
|
| 660 |
|
| 661 |
|
| 662 | if (!ra->ra_pages)
|
| 663 | return;
|
| 664 |
|
| 665 | |
| 666 | * Same bit is used for PG_readahead and PG_reclaim.
|
| 667 |
|
| 668 | if (folio_test_writeback(folio))
|
| 669 | return;
|
| 670 |
|
| 671 | trace_page_cache_async_ra(ractl->mapping->host, index, ra, req_count);
|
| 672 | folio_clear_readahead(folio);
|
| 673 |
|
| 674 | if (blk_cgroup_congested())
|
| 675 | return;
|
| 676 |
|
| 677 | max_pages = ractl_max_pages(ractl, req_count);
|
| 678 | |
| 679 | * It's the expected callback index, assume sequential access.
|
| 680 | * Ramp up sizes, and push forward the readahead window.
|
| 681 |
|
| 682 | expected = round_down(ra->start + ra->size - ra->async_size,
|
| 683 | folio_nr_pages(folio));
|
| 684 | if (index == expected) {
|
| 685 | ra->start += ra->size;
|
| 686 | |
| 687 | * In the case of MADV_HUGEPAGE, the actual size might exceed
|
| 688 | * the readahead window.
|
| 689 |
|
| 690 | ra->size = max(ra->size, get_next_ra_size(ra, max_pages));
|
| 691 | goto readit;
|
| 692 | }
|
| 693 |
|
| 694 | |
| 695 | * Hit a marked folio without valid readahead state.
|
| 696 | * E.g. interleaved reads.
|
| 697 | * Query the pagecache for async_size, which normally equals to
|
| 698 | * readahead size. Ramp it up and use it as the new readahead size.
|
| 699 |
|
| 700 | rcu_read_lock();
|
| 701 | start = page_cache_next_miss(ractl->mapping, index + 1, max_pages);
|
| 702 | rcu_read_unlock();
|
| 703 |
|
| 704 | if (!start || start - index > max_pages)
|
| 705 | return;
|
| 706 |
|
| 707 | ra->start = start;
|
| 708 | ra->size = start - index;
|
| 709 | ra->size += req_count;
|
| 710 | ra->size = get_next_ra_size(ra, max_pages);
|
| 711 | readit:
|
| 712 | ra->order += 2;
|
| 713 | align = 1UL << min(ra->order, ffs(max_pages) - 1);
|
| 714 | end = ra->start + ra->size;
|
| 715 | aligned_end = round_down(end, align);
|
| 716 | if (aligned_end > ra->start)
|
| 717 | ra->size -= end - aligned_end;
|
| 718 | ra->async_size = ra->size;
|
| 719 | ractl->_index = ra->start;
|
| 720 | page_cache_ra_order(ractl, ra);
|
| 721 | }
|
| 722 | EXPORT_SYMBOL_GPL(page_cache_async_ra);
|
| 723 |
|
| 724 | ssize_t ksys_readahead(int fd, loff_t offset, size_t count)
|
| 725 | {
|
| 726 | struct file *file;
|
| 727 | const struct inode *inode;
|
| 728 |
|
| 729 | CLASS(fd, f)(fd);
|
| 730 | if (fd_empty(f))
|
| 731 | return -EBADF;
|
| 732 |
|
| 733 | file = fd_file(f);
|
| 734 | if (!(file->f_mode & FMODE_READ))
|
| 735 | return -EBADF;
|
| 736 |
|
| 737 | |
| 738 | * The readahead() syscall is intended to run only on files
|
| 739 | * that can execute readahead. If readahead is not possible
|
| 740 | * on this file, then we must return -EINVAL.
|
| 741 |
|
| 742 | if (!file->f_mapping)
|
| 743 | return -EINVAL;
|
| 744 | if (!file->f_mapping->a_ops)
|
| 745 | return -EINVAL;
|
| 746 |
|
| 747 | inode = file_inode(file);
|
| 748 | if (!S_ISREG(inode->i_mode) && !S_ISBLK(inode->i_mode))
|
| 749 | return -EINVAL;
|
| 750 | if (IS_ANON_FILE(inode))
|
| 751 | return -EINVAL;
|
| 752 |
|
| 753 | return vfs_fadvise(fd_file(f), offset, count, POSIX_FADV_WILLNEED);
|
| 754 | }
|
| 755 |
|
| 756 | SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
|
| 757 | {
|
| 758 | return ksys_readahead(fd, offset, count);
|
| 759 | }
|
| 760 |
|
| 761 | #if defined(CONFIG_COMPAT) && defined(__ARCH_WANT_COMPAT_READAHEAD)
|
| 762 | COMPAT_SYSCALL_DEFINE4(readahead, int, fd, compat_arg_u64_dual(offset), size_t, count)
|
| 763 | {
|
| 764 | return ksys_readahead(fd, compat_arg_u64_glue(offset), count);
|
| 765 | }
|
| 766 | #endif
|
| 767 |
|
| 768 | |
| 769 | * readahead_expand - Expand a readahead request
|
| 770 | * @ractl: The request to be expanded
|
| 771 | * @new_start: The revised start
|
| 772 | * @new_len: The revised size of the request
|
| 773 | *
|
| 774 | * Attempt to expand a readahead request outwards from the current size to the
|
| 775 | * specified size by inserting locked pages before and after the current window
|
| 776 | * to increase the size to the new window. This may involve the insertion of
|
| 777 | * THPs, in which case the window may get expanded even beyond what was
|
| 778 | * requested.
|
| 779 | *
|
| 780 | * The algorithm will stop if it encounters a conflicting page already in the
|
| 781 | * pagecache and leave a smaller expansion than requested.
|
| 782 | *
|
| 783 | * The caller must check for this by examining the revised @ractl object for a
|
| 784 | * different expansion than was requested.
|
| 785 |
|
| 786 | void readahead_expand(struct readahead_control *ractl,
|
| 787 | loff_t new_start, size_t new_len)
|
| 788 | {
|
| 789 | struct address_space *mapping = ractl->mapping;
|
| 790 | struct file_ra_state *ra = ractl->ra;
|
| 791 | pgoff_t new_index, new_nr_pages;
|
| 792 | gfp_t gfp_mask = readahead_gfp_mask(mapping);
|
| 793 | unsigned long min_nrpages = mapping_min_folio_nrpages(mapping);
|
| 794 | unsigned int min_order = mapping_min_folio_order(mapping);
|
| 795 |
|
| 796 | new_index = new_start / PAGE_SIZE;
|
| 797 | |
| 798 | * Readahead code should have aligned the ractl->_index to
|
| 799 | * min_nrpages before calling readahead aops.
|
| 800 |
|
| 801 | VM_BUG_ON(!IS_ALIGNED(ractl->_index, min_nrpages));
|
| 802 |
|
| 803 |
|
| 804 | while (ractl->_index > new_index) {
|
| 805 | unsigned long index = ractl->_index - 1;
|
| 806 | struct folio *folio = xa_load(&mapping->i_pages, index);
|
| 807 |
|
| 808 | if (folio && !xa_is_value(folio))
|
| 809 | return;
|
| 810 |
|
| 811 | folio = ractl_alloc_folio(ractl, gfp_mask, min_order);
|
| 812 | if (!folio)
|
| 813 | return;
|
| 814 |
|
| 815 | index = mapping_align_index(mapping, index);
|
| 816 | if (filemap_add_folio(mapping, folio, index, gfp_mask) < 0) {
|
| 817 | folio_put(folio);
|
| 818 | return;
|
| 819 | }
|
| 820 | if (unlikely(folio_test_workingset(folio)) &&
|
| 821 | !ractl->_workingset) {
|
| 822 | ractl->_workingset = true;
|
| 823 | psi_memstall_enter(&ractl->_pflags);
|
| 824 | }
|
| 825 | ractl->_nr_pages += min_nrpages;
|
| 826 | ractl->_index = folio->index;
|
| 827 | }
|
| 828 |
|
| 829 | new_len += new_start - readahead_pos(ractl);
|
| 830 | new_nr_pages = DIV_ROUND_UP(new_len, PAGE_SIZE);
|
| 831 |
|
| 832 |
|
| 833 | while (ractl->_nr_pages < new_nr_pages) {
|
| 834 | unsigned long index = ractl->_index + ractl->_nr_pages;
|
| 835 | struct folio *folio = xa_load(&mapping->i_pages, index);
|
| 836 |
|
| 837 | if (folio && !xa_is_value(folio))
|
| 838 | return;
|
| 839 |
|
| 840 | folio = ractl_alloc_folio(ractl, gfp_mask, min_order);
|
| 841 | if (!folio)
|
| 842 | return;
|
| 843 |
|
| 844 | index = mapping_align_index(mapping, index);
|
| 845 | if (filemap_add_folio(mapping, folio, index, gfp_mask) < 0) {
|
| 846 | folio_put(folio);
|
| 847 | return;
|
| 848 | }
|
| 849 | if (unlikely(folio_test_workingset(folio)) &&
|
| 850 | !ractl->_workingset) {
|
| 851 | ractl->_workingset = true;
|
| 852 | psi_memstall_enter(&ractl->_pflags);
|
| 853 | }
|
| 854 | ractl->_nr_pages += min_nrpages;
|
| 855 | if (ra) {
|
| 856 | ra->size += min_nrpages;
|
| 857 | ra->async_size += min_nrpages;
|
| 858 | }
|
| 859 | }
|
| 860 | }
|
| 861 | EXPORT_SYMBOL(readahead_expand);
|